NDA vs data license: why an NDA isn't enough to share data

What an NDA protects, and what it does not#

An NDA protects the secrecy of information shared for a stated purpose, usually evaluating a possible business relationship. It obliges the recipient not to disclose the information and to use it only for that purpose. It rarely says anything about training models, what happens to a model built with the data, or what the discloser is paid.

That design suits business plans and pricing sheets. It fits poorly when the information is a dataset whose main value lies in being used. Once records have shaped a model's weights, returning or destroying the original files does not undo the use, and confidentiality duties are difficult to apply to the model itself.

Mutual NDAs add a further wrinkle. They are drafted for two-way exchange, so their purpose clauses tend to be broad and their exclusions generous, which fits poorly when one party is handing over records and the other is receiving them.

What a data license must add#

A data license must add the terms that turn permission to look into permission to use: what the buyer may do, for how long, in which field, with what protections and for what payment. The table compares typical NDA coverage with what a license for AI use needs.

Not every license needs every term at full strength. An evaluation-only arrangement can be short; a training license for a large archive needs all of these points settled in writing.

Why confidentiality cannot stand in for use limits#

Confidentiality cannot stand in for use limits because a model can carry value from data without ever disclosing it. A recipient could train on records, keep them secret and arguably comply with an NDA whose purpose clause is loose. The discloser's real concern, unapproved use, is not what the NDA was built to control.

Residuals clauses make this worse. Some template NDAs let the recipient use information retained in the unaided memory of its staff. Applied to data, that language invites arguments about what a team learned from the records. Ask to strike or narrow any residuals clause in an NDA that precedes data sharing.

Industry standards point the same way. The Data & Trust Alliance's Data Provenance Standards include a Use group with elements such as license to use and intended data use, reflecting an expectation that a written use basis travels with a dataset rather than being inferred from a confidentiality agreement.

The order of documents from first call to delivery#

The order of documents should track what is being shared at each stage, so protection rises with exposure. Nothing beyond descriptions should move until the right document is signed.

Some parties combine the NDA and evaluation terms into one document. That works if the evaluation limits are explicit; the title on the document matters far less than its terms.

• 1First conversation: no NDA is needed if you share only metadata such as system names, record families and years of history.
• 2Deeper discussion: an NDA once schemas, contract terms or business context are shared.
• 3Sample review: evaluation terms that limit use to assessment, bar training, require deletion and remove any residuals clause.
• 4Full delivery: a signed data license with permitted use, model and output terms, deletion, audit, warranties and payment.
• 5Personal data: a data processing addendum if any personal information remains after preparation.

Mistakes that happen when an NDA is all that exists#

The most damaging mistake is delivering a large export because a buyer asked to start under its standard NDA. Once data has moved without use terms, the supplier's leverage to set them drops sharply, and any later license is negotiated around facts already on the ground.

Other mistakes include accepting a mutual NDA that defines the purpose as exploring any business relationship, leaving out deletion certification, and sharing raw records before privacy preparation. Each is avoidable if the general counsel sets two rules: samples need evaluation terms, and delivery needs a signed license.

A quieter mistake is letting the NDA's term govern the data. NDA confidentiality periods are often limited, while licensed records and anything derived from them may need protection for as long as any copy exists.

Illustrative: a freight audit software company answers a sample request#

Illustrative: a fictional freight audit software company receives a request from an AI developer that wants to begin under its standard mutual NDA and review a large export of support tickets and invoice dispute notes.

The general counsel signs the NDA for discussion only, after striking the residuals clause, and shares a metadata summary of systems, record families and years of history. For the sample, she sends short evaluation terms limiting use to assessment, barring training and requiring deletion with written certification. The sample is prepared first, with carrier and shipper names and contact details removed.

When the developer asks for the full archive, the parties negotiate a license covering fine-tuning and evaluation, deletion of raw files at the end of the term, survival rules for trained weights, an officer certification in place of on-site audit, and payment terms. Delivery happens only after signature.

How SourceX sequences confidentiality and licensing#

SourceX follows the same order of documents. The initial fit check uses metadata only, samples move only after preparation and under evaluation limits, and full delivery happens in the Delivery step of the SourceX five-step transaction, after the supplier's approval.

The resulting SourceX Evidence Packet records provenance, licensing rights, permitted use, the privacy record and release authorization, so the license terms and the delivered package can be checked against each other at any point.